An efficient use of a Distributed Heterogeneous Supercomputing System (DHSS) requhes a thorough understanding of applications and their intelligent scheduling within the system. In this paper we present a general management framework for the DHSS, by introducing an application characterization technique, called Code Flow Graph (CFG) and Code Interaction Graph (CIG). These models are based on code profiling and andyticd benchmarking and provide a detailed archiitecturaldependent characterization of DHSS applications. A generd cost function is presented that is based on the execution and 1/0 overheads associated with applications. An optimd scheduler tries to minimize this cost; the design of which is an NP-complete problem. We describe how network caching can help to reduce the com1)lexity of scheduling in a DHSS.
Heterogeneous Supercomputing, Code Profiling, Benchmarking, Task Flow Graph, Task Intercative Graph, Scheduling, Mapping.
Date of this Version